Add llama and resnet emitc tests #2144
Conversation
svuckovicTT
requested review from
mtopalovicTT,
sdjordjevicTT,
azecevicTT,
jserbedzijaTT and
jnie-TT
as code owners
|
One weird thing here though... in llama test, generating the flatbuffer takes ~75 seconds for some reason, the rest of the compiler is much faster (total 1-2 sec). We should look into what is so slow in the fyi @nsmithtt |
|
@svuckovicTT You might want to try this #1782. |
svuckovicTT
force-pushed
the
svuckovic/add-llama-resnet-emitc-tests
branch
from
e640031 to
fb0972a
Compare
| ArrayAttr arrayAttrs = rewriter.getArrayAttr({ | ||
| rewriter.getIndexAttr(0), // input tensor | ||
| repeatInterleaveOp.getRepeatsAttr(), repeatInterleaveOp.getDimAttr(), | ||
| repeatInterleaveOp.getMemoryConfig().has_value() |
There was a problem hiding this comment.
Nit: has_value() is not needed.
There was a problem hiding this comment.
Sure, but I prefer it this way as it makes it obvious what is being checked.
| repeatInterleaveOp.getMemoryConfigAttr(), | ||
| repeatInterleaveOp.getLoc()) | ||
| ->getResult(0)), | ||
| mlir::cast<Attribute>(rewriter.getIndexAttr(1))) |
There was a problem hiding this comment.
Is this cast really needed?
Also I struggle to follow the logic here, you can maybe derive work related to ?: operator before ArrayAttr creation.
There was a problem hiding this comment.
The cast is indeed needed, otherwise the error is Incompatible operand types ('IntegerAttr' and 'emitc::OpaqueAttr'). I remember that you and I actually looked into this while working on OnesOp conversion haha!
you can maybe derive work related to ?: operator before ArrayAttr creation.
I'll leave it like this since it's "closer" to the automated approach we discussed, with the idea of this op (and OnesOp with same-style implementation) being one of the first to rework into it (automated approach).
- added `assumeVefied()` to `printFlags` that are used for printing op - greately lowers TTNN->Flatbuffer compile time (on LlamaPrefill example from #2144 compile time was lowered from ~70s -> ~1.5s
### Ticket #2156 ### Problem description As described here #1782 `getOpDebugString` was taking more than 99% of the time of `ttmlir-translate -ttnn-to-flatbuffer` execution. From that time, 99% was spent on op verification. ### What's changed - added `assumeVefied()` to `printFlags` that are used for printing ops (the way we are running our pipelines I believe this is a safe assumption) - greatly lowers TTNN->Flatbuffer compile time (on LlamaPrefill example from #2144) compile time was lowered from ~70s -> ~1.5s (around 0.8s was spent on op printing)
svuckovicTT
force-pushed
the
svuckovic/add-llama-resnet-emitc-tests
branch
from
fb0972a to
a07e6b2
Compare
Ticket
Closes #2143
Problem description
Add llama and resnet tests for emitc path.
What's changed
ttnn::repeat_interleavesupport added thru emitc--ttnn-create-input-genswasn't putting tensors on device when they were assumed on device by their layout desc, fixed thatci_compile_dylib.pyto not removettnn-dylib.cppas it is useful for debugging - this is okay as the file is gitignored already